Re: panic: Unaligned free (was: kernel panic while copying files)

From: Gary Jennejohn <gljennjohn_at_gmail.com>
Date: Mon, 5 Jul 2021 18:33:24 +0200
On Mon, 5 Jul 2021 15:04:48 +0100
Edward Tomasz Napiera__a <trasz_at_freebsd.org> wrote:

> On 0701T1330, Gary Jennejohn wrote:
> > Gary Jennejohn <gljennjohn_at_gmail.com> wrote:  
> > > I noticed that the value of vm.debug.divisor affects what value is
> > > returned in uma_core.c:uma_dbg_kskip(), so I decided to try a few
> > > different values.
> > > 
> > > The returned value is used to set skipdbg in uma_core.c:item_dtor().
> > > 
> > > The default is vm.debug.divisor=1.
> > > 
> > > vm.debug.divisor is only present when INVARIANTS is defined.
> > > 
> > > kskipdbg eventually affects the value of freei.
> > > 
> > > With these values:
> > > vm.debug.divisor: 0
> > > kern.cam.da.enable_uma_ccbs: 1
> > > I can turn on the disk and it comes up without a panic!
> > > 
> > > However, I didn't try to do any large data transfers to the disk.
> > > 
> > > So, it appears that at least vm.debug.divisor is a big factor in
> > > whether or not a panic happens with INVARIANTS.
> > >   
> > 
> > I decided to do a real test.  So I built a kernel w/o INVARIANTS and
> > installed it to /boot/test.
> > 
> > Then I stuck a 160GB disk I had around into an external USB3 enclosure
> > and put a filesystem on it.
> > 
> > The I booted the new kernel from /boot/test and set the sysctls so:
> > kern.cam.da.enable_uma_ccbs: 1
> > kern.cam.ada.enable_uma_ccbs: 1
> > 
> > After that I plugged in the external USB3 enclosure and copied about
> > 114GiB of data from an internal SSD to it - without a kernel panic:
> > Filesystem    Size    Used   Avail Capacity  Mounted on
> > /dev/da0p1    144G    114G     18G    86%    /mnt
> > 
> > I'm pretty sure that's more than I could copy without a kernel panic
> > prior to the recent changes made in cam and umass.
> > 
> > My test may not be real proof that all bugs have been squashed, but it
> > certainly seems to be a better situation than we had before.  
> 
> I think the vm.debug.divisor simply masks the problem; the underlying
> bug is still there.
> 
> Could you go back to the setup which panics, and then test the patch
> at https://reviews.freebsd.org/D31054?  It fixes the scenario described
> by Warner.
> 

It looks like this patch fixes things.

I used the default value vm.debug.divisor=1 and both enable_uma_ccbs=1
(which are now the default values on my system).

I used the 8TiB disk, which spins up very slowly and usually resulted very
quickly in a panic - no panic with the patch.

Then using dd to /dev/null (bs=1m) I transferred:

308755+0 records in
308755+0 records out
323753082880 bytes transferred in 1366.162410 secs (236979938 bytes/sec)

from the disk, so about 324GiB without a panic.

-- 
Gary Jennejohn
Received on Mon Jul 05 2021 - 16:33:24 UTC

Original text of this message